Reflex retweetledi
We released Reflex 0.9 - every action in your app is now callable by an AI agent out of the box.
We ran a benchmark comparing two approaches to letting AI agents interact with a web app:
- A vision agent (browser/computer use) that screenshots the UI and clicks around
- An API agent that calls HTTP endpoints directly
The task for both agents was to find a "Smith" customer with the most orders, accept their pending reviews, and mark their most recent order as delivered. We chose this task since it's similar to automation work a typical tool sees.
The vision agent took 550k tokens and 17 minutes on average. The API agent took 12k tokens and 19.7 seconds. Of course, API agents are faster and more token efficient since they don't need to take screenshots and directly interface with the UI. The problem is many apps don't have APIs for every action, since it takes engineering overhead to build and maintain each separate API codebase.
We built a plugin for Reflex that auto-generates HTTP endpoints from your app's existing event handlers. Your app has a button with an on_click handler, the plugin exposes that handler as an endpoint. An agent can call the same function a human click triggers. No separate API to build or maintain.
As agents become a bigger share of your app's users, every action needs to be callable by code - not just by clicking a button.
Benchmark: reflex.dev/blog/vision-ag…
English
